Prev: PEEEEEEP
Next: Texture units as a general function
From: Robert Myers on 3 Jan 2010 06:25 On Jan 3, 2:59 am, Mayan Moudgill <ma...(a)bestweb.net> wrote: > Andrew Reilly wrote: > > On Sat, 02 Jan 2010 19:13:57 -0800, Robert Myers wrote: > > >>It's conceivable that I could find a > >>combination of @Home settings and Computer Power Management setting that > >>would result in negligible extra power, but how many will do that? > > > I'm fairly confident that Mayan's use of @Home as a tag was more along > > the lines of "computations that are known to require lots of grinding on > > small chunks of data, and so therefore trivially parallelizable with > > arbitrary network geometry", rather than the "free cycles" aspect. > > Precisely! > > Robert has a (probably legitimate) persistent complaint about the > inability to solve certain classes of physics problems because of > computational resources. > > So: are the physics problems that are being tackled on a BlueGene-style > supercomputer at LLNL do-able on a CoW/NoW? If not, what makes it > possible to do it only on a BlueGene-style machine? > Blue Gene had three attractive properties: low power consumption, computational density, and low network latency. I suspect that all three were requirements for the original LLNL acquisition. The one that's hardest to meet (given new hardware) is network latency. Some fairly clever engineering of heat transfer made the computational density of Blue Gene possible, but advances in hardware since then probably make it easy to achieve the same computational density without heroic measures. Similarly with power consumption. If you can achieve the required network latency with available commercial solutions, then you don't need a Blue Gene, and you will have much more control over the network properties, including bi-section bandwidth. > [Robert also claims that the results based on access to supercomputers > are externally unverifiable, suspect, and possibly incorrect. The first > may be true, and the rest do follow as conclusions. However, I'm not > interested in tackling that subject. I just mention it so that anyone > who doesn't want to address the technical question can have something to > quote when they post a response.] I've made myself tedious by complaining that more linpack flops <> progress. I don't regret it at all, because either I or someone else has been heard. I don't care who is heard, as long as someone has been. In considering the prospect of billions of processors, I think it's fair to ask: why? I've identified one reason why not. It's hard to get time on the really big/good telescopes, too, but, short of actual manipulations of data or images, you sort of have to take what the stars give you, and the chances of getting caught fudging are very high. Not so with computations that take forever on computers of which perhaps only one or two in the world exist. My only suggestion is that we think hard about the wisdom of letting science be driven by such machines. We can, thanks to Moore's law, do an awful lot of physics with fairly modest resources, and it might be a wiser use of resources to make mid-scale machines more available and more usable than concentrating so much on "super-computers," especially since the desktop I'm working on would make the Cray-I look pathetic by comparison. To the extent that comp.arch has been misused as a PR platform for bigger is better, and it has been, I think it only fair that there should be a dissenting voice. Robert.
From: nmm1 on 3 Jan 2010 06:44 In article <7qacstFk5U1(a)mid.individual.net>, Del Cecchi <delcecchi(a)gmail.com> wrote: > >Since I've argued that super-sized computers seem to me to be of >questionable value, maybe that's all the vast bulk of science really >needs. If I really need a computer with respectable bi-section >bandwidth, I can skip waiting for a gigantic machine that runs at 5% >efficiency (or worse) and learn to live with whatever I can build >myself. Bisection bandwidth is the Linpack measurement of networking, but that doesn't affect your point. I favoured (and favour) the latency and bandwidth of all-to-all, both because it is more realistic and because it is invariant over the network topology and implementation. The only real point of the specialist HPC systems nowadays is when the problems are limited by communication and not processing, but it's more a situation of how you spend your money than differences in architecture. There isn't any major technical problem with using multiple InfiniBand ports on each node, and a suitable topology; the software can get interesting, though. However, people who do that usually want SMALLER nodes, because that increases the network capacity relative to computation power. It's expensive to scale the former pro-rata to the latter, and is often not feasible with off-the-shelf components. I.e. they use single- socket nodes. Regards, Nick Maclaren.
From: Mayan Moudgill on 3 Jan 2010 10:03 Robert Myers wrote: > On Jan 3, 2:59 am, Mayan Moudgill <ma...(a)bestweb.net> wrote: > >>So: are the physics problems that are being tackled on a BlueGene-style >>supercomputer at LLNL do-able on a CoW/NoW? If not, what makes it >>possible to do it only on a BlueGene-style machine? >> > > Blue Gene had three attractive properties: low power consumption, > computational density, and low network latency. I suspect that all > three were requirements for the original LLNL acquisition. The one > that's hardest to meet (given new hardware) is network latency. Some > fairly clever engineering of heat transfer made the computational > density of Blue Gene possible, but advances in hardware since then > probably make it easy to achieve the same computational density > without heroic measures. Similarly with power consumption. If you > can achieve the required network latency with available commercial > solutions, then you don't need a Blue Gene, and you will have much > more control over the network properties, including bi-section > bandwidth. > > When I asked for a physics problem that was being tackled on BlueGene that couldn't be done on a CoW/NoW, I hadn't considered the possibility that BlueGene itself was the physics experiment! I'm assuming (based on the reply) the scientists are using the spot heating due to the CPU and the resulting heat transfer to mirror some real world physics problem. Or did I miss the point entirely?
From: Mayan Moudgill on 3 Jan 2010 10:29 Rob Warnock wrote: > Mayan Moudgill <mayan(a)bestweb.net> wrote: > +--------------- > | Rob Warnock wrote: > | > Robert Myers <rbmyersusa(a)gmail.com> wrote: > | > +--------------- > | > | I doubt if operating systems will ever be written in an elegant, > | > | transparent, programmer-friendly language. > | > +--------------- > | > > | > So the various Lisp Machines never existed? ;-} ;-} > | > | Exactly how is LISP elegant, transparent & programmer-friendly? > +--------------- > > See <http://www.paulgraham.com/avg.html>... > Why do you think that this proves that LISP is elegant, transparent & programmer-friendly? The fact that it took 2+ people 2+ years to write a 25,000 line LISP application (Viaweb which became Yahoo!Store)? see: http://lib.store.yahoo.net/lib/paulgraham/bbnexcerpts.txt [And while we're talking programmer friendliness - after being acquired by Yahoo!, Viaweb was rewritten in C++ - probably for performance] If you need performance, particularily in an application such as an OS, I suspect that the resulting LISP will be laden with destructive operations & programmer-managed memory. (Just for context - I've used Symbolics Machines & MacIvories, about 20! years ago, to write at least a couple of decent sized apps) (This was the same job where the only editor available was emacs. It was a relief when it was over, and I could to escape to vi! The good thing about that was that I can now tell people that I had to use emacs exclusively for most of a year, and I still prefer vi, so my preference is actually experimental and not pure pre-judgement, and don't they think they should try vim for a few months before claiming emacs is better, hmmm?)
From: nmm1 on 3 Jan 2010 10:52
In article <dtKdnWNf3IldLN3WnZ2dnUVZ_gGdnZ2d(a)bestweb.net>, Mayan Moudgill <mayan(a)bestweb.net> wrote: > >When I asked for a physics problem that was being tackled on BlueGene >that couldn't be done on a CoW/NoW, I hadn't considered the possibility >that BlueGene itself was the physics experiment! I'm assuming (based on >the reply) the scientists are using the spot heating due to the CPU and >the resulting heat transfer to mirror some real world physics problem. The CETEP/CASTEP/ONETEP series of ab initio quantum mechanics programs alternate between the solution of a large set of simultaneous equations and a very large 3-D FFT. The latter, like sorting, is limited by communication. It is (amongst other things) used for surface chemistry caclulations, which are of immense industrial importance. There are dozens of other examples - that's just one I happen to know a little about. Regards, Nick Maclaren. |